Review Rarely Beats the Stronger Model: Separating Verification from Capability Transfer in LLM Pipelines

Yirui Zhang; Yash Raj Shrestha

cs.CLOct 11, 2026

Review Rarely Beats the Stronger Model: Separating Verification from Capability Transfer in LLM Pipelines

Authors: Yirui Zhang, Yash Raj Shrestha

Organizations: University of Lausanne

Abstract

Qwen3-8B reviewing Llama-3.1-8B in math adds 33.2 points, yet Qwen3-8B alone is 6.4 points more accurate than the result. Reviewing another model's answer is usually judged by its gain over the generator. We show that this confuses verification, which adds accuracy beyond either model alone, with capability transfer, where a stronger reviewer supplies answers it could have given on its own. Measured against the stronger model alone, across 48 pairs of 12 open-weight models (3B to 72B) in math, knowledge, code and reasoning, review beats it in only 9 of 96 review directions, and does worse the larger the capability gap. The reason is anchoring: the reviewer mostly follows the generator, discarding its own correct answers. Review costs 2.6 to 2.8 times as much as the stronger model alone, while a selector using the same two calls is 2.7 to 4.2 points more accurate than review and cheaper. Studies of verification should report the stronger model alone, not only the gain over the generator.

Figures & tables

Figure 1: Overview. (a) In review, the reviewer (V) reads the generator’s (G) answer and returns the final answer; in selection, a selector picks one of the two models’ independent answers. Both use two calls. We compare the final answer with G alone (

\Delta_{G}

) and with the stronger model alone (

\Delta_{S}

). (b) Outcomes of the 48 directions of each kind. When the stronger model reviews, review beats G in 42 but the stronger model in only one; when the weaker one reviews (G is stronger), it beats both in 8 and is worse than G in 22. (c) After review, the final answer differs from G’s in correctness far less often than V’s own answer did. (d) With the same two calls, choosing between the two independent answers is more accurate than review; the oracle picks a correct answer when either model is correct.

Figure 2: A reviewer that solves a question alone follows the generator’s wrong answer after reading it. Llama-3.1-8B lists primes only up to 29 and answers 10. Qwen3-8B answers 11 (correct) on its own; reviewing, it writes the correct bound but copies the list that stops at 29, and answers 10.

Work	Second model’s role	Final answer	Reference (as defined in the paper)	Stronger model alone	Equal-call choice	Split by who is right alone
Huang et al. (2024)	Same model revises its own answer	Same model	Accuracy after vs. before self-correction	n/a	✓	n/a
Singhi et al. (2025)	Same model scores its own solutions	G’s sample	Success rate vs. self-consistency at equal compute	n/a	✓	n/a
Lifshitz et al. (2025)	Verifiers approve G’s samples; most approved is kept	G’s sample	Accuracy of BoN-MAV vs. pass@1 of G	when G is stronger	✓	$\times$
Lu et al. (2026)	Verifier accepts or rejects; G resamples	G’s sample	$\mathrm{Precision}(S,V)-\mathrm{SolverAcc}(S)$	$\times$	$\times$	$\times$
Zhou et al. (2026)	Verifier filters G’s samples	G’s sample	$\Delta p_{V}=p_{G,V}-p_{G}$	one case	$\times$	$\times$
Xiang et al. (2026)	Other model reviews and rewrites	Reviewer	Pass rate vs. each model alone	one pair	$\times$	$\times$

Table 1: Closest prior work. Most studies let the second model choose among the generator’s own samples and measure the gain over the generator; in review, the second model writes the final answer. “Equal-call choice”: a baseline that picks among independent answers at the same cost. “Split by who is right alone”: outcomes split by whether each model answers correctly on its own.

Domain	Source	$N$	Grading
Math	MATH, levels 3–5	2,500	symbolic equiv.
Knowledge	MMLU-Pro	2,500	option letter
Code	BigCodeBench-Instruct	832	unit tests
Reasoning	ReClor, FOLIO v2, GPQA	901	option letter

Table 2: Tasks, from MATH ( Hendrycks et al., 2021 ) , MMLU-Pro ( Wang et al., 2024b ) , BigCodeBench ( Zhuo et al., 2025 ) , and for reasoning 500 ReClor ( Yu et al., 2020 ) , 203 FOLIO v2 ( Han et al., 2024 ) and 198 GPQA Diamond ( Rein et al., 2024 ) questions.

Figure 3: Each point is one review direction, coloured by who reviews whom; shape gives the domain, and hollow markers have a 95% interval containing zero. (a)

\Delta_{G}

against the gap; on the dashed line the final answer matches the stronger model alone. (b)

\Delta_{S}

against the gap. (c) The gap of every direction at each outcome level; numbers count the directions of each colour.

Figure 4: Why review leaves the second model’s value unused. Each dot is one review direction, coloured as in Figure 3 ; black bars are means with domains weighted equally. (a) Accuracy relative to the stronger model alone of the oracle of the two solo answers (G or V right), the oracle after review (G’s answer or the final answer right), and the final answer. (b) On questions only one model gets right alone, how often the final answer is right. (c) The change over G alone, by which model is right alone. (d) Weaker model reviews: on questions only G gets right alone, how often the final answer stays right, against

\Delta_{S}

; colours give the outcome level.

Figure 5: The cost and accuracy of using a second model; cost per question is parameters

\times

(input + output tokens). (a) Each way of using the two models relative to the stronger model alone: means over 48 pairs with domains weighted equally, with 95% intervals; JEV also pays an API fee not on this axis. (b) The selector against each review direction, one point per pair; hollow points have a 95% interval that includes zero.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Cell	Case	Pair used
Ties and small tiers
Knowledge, strong-middle, most	Two pairs tie	The pair already run on the development questions
Knowledge, strong-weak, fewest	Two pairs tie	The pair whose weaker model is weaker
Reasoning, middle-middle	The middle tier has two models	Its one pair
Reasoning, strong-middle, fewest	Two pairs tie	Both (12 pairs in all)
Knowledge, strong-strong, fewest	Only three strong models; without Qwen2.5-72B one pair is left	Qwen2.5-72B with Mistral-Small-24B, kept

Appendix

Table 3: Pairs chosen by a further rule. Numbers in parentheses: development questions that only the weaker model answers correctly.

Domain	Tiers	Headroom	Stronger	Acc.	Weaker	Acc.	Gap	$h$
Math	strong-strong	most	Qwen3-8B	80.3	Qwen3-4B	78.3	2.0	6.3
	strong-strong	fewest	Qwen3-32B	85.1	Qwen3-8B	80.6	4.5	4.2
	strong-middle	most	Qwen3-4B	78.3	Mistral-Small-24B	65.4	12.9	5.4
	strong-middle	fewest	Qwen3-32B	85.1	Qwen2.5-3B	59.2	25.9	1.7
	strong-weak	most	Qwen3-8B	80.6	Ministral-8B	50.8	29.7	2.3
	strong-weak	fewest	Qwen3-8B	80.6	Llama-3.1-8B	40.9	39.6	1.2

Appendix

Table 4: The 48 model pairs. Tiers and the headroom end (most or fewest questions that only the weaker model answers correctly) come from the development questions; accuracy, gap and headroom

h

(%) are over all questions used in the analyses (2,500 in math and knowledge, 832 in code, 901 in reasoning). In math, the Qwen3-4B and Qwen3-8B pair uses their answers from an earlier run on the test questions, so Qwen3-8B’s accuracy differs by 0.2 points from its other pairs.

	Math	Knowledge and reasoning	Code
Each answer
Basic	7: mean log-probability and largest logit over the output and over math expressions; answer parseable; question and output length in characters	22: answer parseable; number of options; question length; prompt and output tokens; truncated; 12 token statistics over the output; log-probability, entropy, margin and largest logit at the answer letter	17: question length; prompt and output tokens; truncated; number of code blocks; 12 token statistics over the output
Token spans	33: 11 token statistics each over the output, the math expressions and the boxed answer	–	22: 11 token statistics each over the code and the prose
Content	4: output tokens; truncated; number of boxed answers; self-correction words	16: spread of the answer-letter probability over the options (7); letters mentioned, answer lines, whether the first letter is the last, hedging, elimination and self-correction words (6); digits, numbers and equals signs (3)	26: program structure from the syntax tree (16); use of the required libraries and of the exceptions the task asks for (3); undefined names, unseeded randomness, prints, main guard, file access (5); output characters and self-correction words (2)
Total	44	38	65
Question	8: difficulty level; 7 subject indicators	18: 14 subject and 4 source indicators (zero in reasoning)	3: required libraries; whether and how many exceptions the task asks for

Appendix

Table 5: Selector features per domain, with counts. The logistic regression sees one answer’s features and the question features; the tree sees both answers, the question and the pair. Token statistics are the mean, minimum, 10th percentile and standard deviation of the log-probability of each output token, the mean, 90th percentile and maximum entropy, the mean and 10th percentile of the margin between the two most likely tokens, and the share and longest run of tokens with probability below 0.5; the 12 statistics add the mean largest logit.

	Main	Output $\times 4$					Pairs cheaper
	All	All	Math	Knowl.	Code	Reas.	Main	$\times 4$
Relative to the stronger model alone
Stronger reviews	2.76	2.24	2.17	1.92	3.05	1.83
Weaker reviews	2.56	2.05	2.03	1.75	2.66	1.75
Selector	1.93	1.94	2.12	1.77	2.11	1.77
Selector relative to review

Appendix

Table 6: Cost ratios under the main cost

c=N\,(t_{\text{in}}+t_{\text{out}})

and with output tokens weighted four times,

c^{\prime}=N\,(t_{\text{in}}+4\,t_{\text{out}})

: the ratio is taken per pair, then averaged with domains weighted equally. Pairs cheaper: of 48, the pairs where the selector costs less than review.

Both kinds of direction pooled
	48 pairs	Leaving out one model	All completed pairs
	(96 directions)	(range of the mean)	(131 directions)
Weaker reviews $-$ stronger reviews	$+1.5$ [ $+1.1$ , $+1.9$ ]	$+1.2$ to $+1.9$	$+1.7$ [ $+1.4$ , $+2.1$ ]
$\Delta_{G}$	$+3.0$ [ $+2.7$ , $+3.3$ ]	$+2.5$ to $+3.5$	$+3.2$ [ $+3.0$ , $+3.5$ ]
$\Delta_{S}$	$-2.3$ [ $-2.6$ , $-2.0$ ]	$-2.7$ to $-1.6$	$-2.5$ [ $-2.8$ , $-2.2$ ]
Net change where both agree	$+0.2$ [ $+0.0$ , $+0.4$ ]	$+0.1$ to $+0.4$	$+0.2$ [ $+0.0$ , $+0.4$ ]

Appendix

Table 9: Robustness of the main averages (points, domains weighted equally, 95% intervals that resample questions). Leaving out one model removes every direction involving it; the column gives the lowest and highest mean over the 12 models. All completed pairs adds every other pair we ran review on; it is computed with both kinds of direction pooled only.

	Mean	Questions	Pairs and questions
$\Delta_{S}$ , stronger reviews	$-3.1$	[ $-3.5$ , $-2.7$ ]	[ $-4.1$ , $-2.0$ ]
$\Delta_{S}$ , weaker reviews	$-1.6$	[ $-1.8$ , $-1.3$ ]	[ $-2.4$ , $-0.8$ ]
Oracle over the stronger model	$+7.5$	[ $+7.2$ , $+7.8$ ]	[ $+6.7$ , $+8.3$ ]
Oracle after review, stronger reviews	$+0.6$	[ $+0.2$ , $+1.0$ ]	[ $-0.5$ , $+1.6$ ]
Oracle after review, weaker reviews	$+2.8$	[ $+2.6$ , $+2.9$ ]	[ $+2.3$ , $+3.3$ ]
Net change where both agree, stronger reviews	$+0.9$	[ $+0.7$ , $+1.2$ ]	[ $+0.5$ , $+1.4$ ]

Appendix

Table 10: Main averages (points, domains weighted equally) with 95% intervals that resample questions only, as in the main text, or model pairs and questions.

G	V	Final	What happens	Math	Know.	Code	Reas.	All	Math	Know.	Code	Reas.	All
				Stronger model reviews					Weaker model reviews
✓	✓	✓	Both right alone; review keeps it.	49.5	39.9	24.6	47.5	40.4	49.4	40.5	23.8	48.1	40.4
✓	✓	$\times$	Both right alone; review breaks it. Harm only review can do: a choice between the two answers cannot lose these.	0.9	2.0	2.3	2.6	1.9	1.0	1.4	3.1	2.1	1.9
✓	$\times$	✓	Only G right; review keeps G’s answer.	4.7	6.1	6.3	6.0	5.8	18.6	17.0	10.4	16.8	15.7
✓	$\times$	$\times$	Only G right; review loses it (the G-only part of $\Delta_{G}$ ).	0.7	2.6	1.4	2.5	1.8	1.6	2.5	3.4	2.4	2.5
$\times$	✓	✓	Only V right; review takes up V’s answer (the V-only part of $\Delta_{G}$ ). Capability transfer: V alone answers these right too.	9.5	10.2	4.0	9.9	8.4	0.5	2.3	1.1	1.7	1.4

Appendix

Table 11: All eight cells: the share of questions (%) by whether G alone, V alone and the final answer are right, averaged over the 12 directions of each kind in a domain; All weights the domains equally. The three parts of

\Delta_{G}

in Section 3.1 are the V-only row, minus the G-only row, plus the both-wrong-then-right row minus the both-right-then-wrong row. Columns may not sum to 100 because of rounding.

Feature	Odds ratio	Kept (%), Q1 to Q4	Odds ratio	Kept (%), Q1 to Q4
	Stronger model reviews (15,364)		Weaker model reviews (5,883)
V’s solo output length	0.75 [0.71, 0.79]	54 / 51 / 47 / 39	0.76 [0.69, 0.83]	23 / 21 / 19 / 15
V’s mean maximum logit	1.16 [1.10, 1.22]	43 / 46 / 51 / 52	1.24 [1.11, 1.38]	16 / 18 / 21 / 23
V’s mean log-probability	1.06 [1.00, 1.13]	47 / 48 / 48 / 48	1.04 [0.93, 1.18]	18 / 17 / 20 / 21
G’s output length	1.00 [0.93, 1.07]	52 / 49 / 46 / 45	0.94 [0.84, 1.06]	21 / 20 / 19 / 18
G’s mean maximum logit	0.96 [0.91, 1.01]	48 / 47 / 49 / 48	0.96 [0.87, 1.06]	20 / 17 / 19 / 21

Appendix

Table 12: When the reviewer keeps its own right answer, on questions where G was wrong and V right alone (number of questions in parentheses). Odds ratio: change in the odds that the final answer stays right per standard deviation of the feature, from one logistic regression per kind of direction with all seven features and direction fixed effects; features are standardised within each direction, since raw logits differ in scale across models; 95% intervals clustered by question; bold intervals exclude 1. Kept: share of these questions whose final answer stays right, by quartile of the feature within each direction (Q1 lowest). Correlations, not causal effects.

Figure 6: Where review lands, by who reviews whom and the gap. Each bar is one gap range, split by outcome level (paired 95% intervals); numbers count directions. (a) The stronger model reviews the weaker. (b) The weaker reviews the stronger; G is then the stronger model, so only three levels occur.

Figure 7: After review, the final answer agrees with G more often. Each point is one review direction. x: how often the reviewer answering alone differs from G in correctness; y: how often the final answer does. Points below the dashed line moved toward G; crosses are means with domains weighted equally.

Explore similar work

Reviewer Capability Governs Rejection Targeting, Not Repair Skill: Evidence from LLM Execute-Review-Revise Pipelines
Sep 2, 2026Faizan TanveerLLM Answer Verification LLM Agent Reliability
National University of Computer and Emerging Sciences (FAST NUCES)
Accuracy Hides How Language Models Fail: Measuring Failure States Under Matched Output Budgets
Jul 27, 2026Zongyou Yang, Yinghan HouLLM Evaluation LLM Answer Verification
Dyson School of Design Engineering Imperial College London London, United Kingdom · Department of Electrical and Electronic Engineering Imperial College London London, United Kingdom
Do We Need Frontier Models to Verify Mathematical Proofs?
Apr 2, 2026Aaditya Naik, Guruprerana Shabadi, Rajeev Alur +1Reasoning Verification LLM Reasoning
University of Pennsylvania, U.S.A.

Sep 2, 2026J/K move · Enter open · S save

Reviewer Capability Governs Rejection Targeting, Not Repair Skill: Evidence from LLM Execute-Review-Revise Pipelines

Faizan Tanveer

National University of Computer and Emerging Sciences (FAST NUCES)

LLM Answer Verification LLM Agent Reliability LLM Self-Correction LLM-as-a-Judge

Open paper AskPDF arXiv

Multi-agent LLM pipelines increasingly assign roles, including execution and verification, to models of different capability tiers. This is done because running a flagship model at every stage is expensive. Previous literature has established that verification stages are not always beneficial, but holds reviewer capability roughly fixed relative to the executor. We vary it. We replace the reviewer with models spanning a capability range down to one that cannot solve the problems at all, and measure the outcome of every individual rejection. This is done across a constant set of 100 olympiad mathematics problems. A cross-family mid-tier reviewer improves final accuracy by 12 percentage points, from 52 to 64 percent (p = 0.0005), with zero damaged answers. Same-model self-review attains the highest error-detection rate of any condition (0.85 recall) yet yields no significant gain: it rejects 2.1 times as often for a third the repair rate (15 against 43 percent, p = 0.0074) and falsely rejects 35 percent of its own correct answers against 2 percent for the cross-family reviewer (paired p = 0.000015). The low damage rate of self-review proves to be an artifact of revision inertia rather than reviewer quality: of 18 falsely rejected correct answers, the three where the executor complied all became wrong, while the fifteen it ignored survived unchanged. Below a capability floor the role becomes inert: our weakest reviewer changed zero of 100 final answers while doubling token cost. These findings describe a single executor-reviewer configuration on 100 problems and should be read as a controlled pilot rather than a general claim about verification stages.

Abstract

Explore similar work

Reviewer Capability Governs Rejection Targeting, Not Repair Skill: Evidence from LLM Execute-Review-Revise Pipelines

Accuracy Hides How Language Models Fail: Measuring Failure States Under Matched Output Budgets

Do We Need Frontier Models to Verify Mathematical Proofs?